Object matching method based on attribute alignment and feature fusion

Abstract

The embodiment of the invention provides an object matching method based on attribute alignment and feature fusion, and the method comprises the steps: obtaining attribute description text data of a first target object, carrying out the feature extraction of the attribute description text data, and obtaining a plurality of first attribute description feature phrases; matching the plurality of first attribute description feature phrases with a set attribute classification table to cluster the first attribute description feature phrases to obtain a plurality of first-dimension attribute feature sequences for describing the first target object; vectorizing each first-dimension attribute feature sequence to obtain a corresponding first attribute feature vector so as to form a feature portrait of the first target object; acquiring attribute description text data of a second target object, and performing feature extraction on the attribute description text data to obtain a plurality of second attribute description feature phrases to form a feature portrait of the second target object so as to judge whether the first target object is matched with the second target object or not.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to an object matching method based on attribute alignment and feature fusion. Background Art

[0002] With the development of the Internet, various business models based on online platforms continue to emerge. As an important resource, the company's Internet celebrity database plays a key role in many business scenarios such as marketing promotion and content dissemination. In order to achieve efficient use of Internet celebrity resources and accurately recommend Internet celebrities to suitable partners (such as brands, event organizers, etc.), recommendation algorithms have become the core technical support.

[0003] In the existing recommendation algorithm field, collaborative filtering recommendation algorithm is widely used, which is mainly divided into two common modes: user-based recommendation and item-based recommendation. For example, in user-based recommendation, when it is found that user A likes product P1, and user B and user A show similarities in some aspects, product P1 will be recommended to user B; while in item-based recommendation, if user A likes product P1, and product P2 has similar features to product P1, then product P2 will be recommended to user A.

[0004] However, no matter which collaborative filtering mode is used, in the key link of calculating the similarity of objects, traditional similarity algorithms such as dot product function, cosine theorem or Pearson coefficient are generally used. These algorithms have significant limitations in the application process, and they often require that the data involved in the calculation is a numeric field. Taking the recommendation scenario of the Internet celebrity library as an example, if you want to determine whether the Internet celebrity matches the potential cooperation object (such as the product corresponding to the brand), for the product, you may be able to use numerical parameters such as price (such as both are around 100 yuan), category (such as both are category 1), user rating (such as both are around 80 points), etc., according to the corresponding formula (such as the cosine theorem, whose value range is [-1, 1], the closer to 1, the more similar) to calculate the value representing the similarity.

[0005] However, in the actual Internet celebrity database scenario, the attribute information of the object is rich and varied. In addition to the above-mentioned quantifiable numerical fields, it also contains a large number of non-numerical fields, such as the content style of the Internet celebrity (such as funny, literary, etc.), the field of expertise (such as food, beauty, etc.) and the cooperation preference (such as preferring long-term cooperation or short-term cooperation, etc.). These non-numerical fields are crucial for comprehensively and accurately measuring the matching degree between the Internet celebrity and the cooperation object. However, when traditional similarity algorithms process these non-numerical fields, complex preprocessing operations are required, such as encoding non-numerical attributes into numerical form, which not only increases the complexity and workload of data processing, but also may lose some key semantic information during the conversion process, thereby affecting the accuracy of similarity calculation and the final recommendation effect.

[0006] In addition, relying solely on limited numerical attributes to measure the degree of match makes it difficult to comprehensively characterize the characteristics of influencers and partners from multiple dimensions. As a result, the recommendation results can only reflect the similarity of some surface features and cannot deeply explore the deep-level adaptation relationship, which greatly reduces the accuracy and effectiveness of the recommendations and makes it difficult to meet the high requirements for precise matching recommendations in actual business. Summary of the invention

[0007] In view of this, an embodiment of the present invention provides an object matching method based on attribute alignment and feature fusion to at least partially solve the above problems.

[0008] According to a first aspect of an embodiment of the present invention, there is provided an object matching method based on attribute alignment and feature fusion, which includes:

[0009] Acquire attribute description text data of the first target object, and perform feature extraction on the text data to obtain a plurality of first attribute description feature phrases;

[0010] Using a plurality of the first attribute description feature phrases to match with a set attribute classification table to cluster the first attribute description feature phrases to obtain a plurality of first dimension attribute feature sequences describing the first target object;

[0011] Vectorizing each first-dimensional attribute feature sequence to obtain a corresponding first attribute feature vector to form a feature profile of the first target object;

[0012] Acquire attribute description text data of the second target object, and perform feature extraction on the text data to obtain a plurality of second attribute description feature phrases;

[0013] Aligning a plurality of the second attribute description feature phrases with a plurality of the first attribute description feature phrases to cluster the second attribute description feature phrases to obtain a plurality of second dimension attribute feature sequences describing the second target object;

[0014] Vectorizing each second-dimensional attribute feature sequence to obtain a corresponding second attribute feature vector to form a feature profile of the second target object;

[0015] Based on the feature portrait of the first target object and the feature portrait of the second target object, the matching degree between the first target object and the second target object is calculated and compared with the set matching degree threshold to determine whether the first target object matches the second target object.

[0016] In the solution of the embodiment of the present invention, the object matching method based on attribute alignment and feature fusion has the following technical advantages:

[0017] (1) Effectively process non-numeric fields and reduce data processing complexity

[0018] The method provided by the embodiment of the present invention starts with obtaining the attribute description text data of the object, and directly extracts the features of the text data to obtain the attribute description feature phrases. For example, for the text description information of the Internet celebrity such as "funny style, good at food, inclined to long-term cooperation", the corresponding feature phrases can be naturally extracted. Then, clustering is performed by matching with the set attribute classification table to form attribute feature sequences of different dimensions. The whole process does not require the non-numeric fields to be deliberately converted into numerical forms, and cleverly avoids the complex preprocessing links in traditional methods that are prone to information loss, effectively reducing the complexity of data processing, ensuring the integrity of the original semantic information of the data, and laying the foundation for subsequent accurate matching calculations.

[0019] (2) Comprehensive integration of multi-dimensional features to improve the comprehensiveness of matching measurement

[0020] The method of the present invention fully considers the rich attribute information of the object. Whether it is a numerical field or a non-numerical field, it undergoes a series of steps such as feature extraction, clustering, and vectorization to finally form a complete feature portrait. Taking the cooperation and matching between Internet celebrities and brand merchants as an example, the feature portrait of the Internet celebrity integrates the vectors corresponding to multi-dimensional features such as their content creation style and areas of expertise. The brand merchant also integrates multi-dimensional feature vectors such as its own product characteristics and brand positioning. Based on such a feature portrait that fully integrates multi-dimensional features to calculate the matching degree, it is possible to deeply explore the deep-level adaptation relationship between objects, and comprehensively consider whether the two match from multiple angles, which greatly improves the comprehensiveness of the matching degree measurement, and is no longer limited to the comparison of surface numerical attributes, so that the recommendation results are more in line with the high requirements for precise matching in actual business.

[0021] (3) Improve the accuracy of similarity calculation and optimize the recommendation effect

[0022] The method proposed in the present invention constructs a feature portrait through attribute alignment and feature fusion, ensuring the accurate extraction and reasonable integration of feature information in each dimension. When calculating the matching degree between the first target object and the second target object, it is based on the feature portrait that completely and accurately reflects the characteristics of the object. For example, when matching Internet celebrities with potential cooperative products, all kinds of key information contained in the feature portraits of both parties can be fully reflected in the matching degree calculation process, avoiding calculation errors caused by information loss or one-sided considerations, making similarity calculations more accurate, so that objects with truly high matching degrees can be accurately screened for recommendation, significantly optimizing the recommendation effect, and helping to better realize the efficient use of Internet celebrity resources, and meet the company's Internet celebrity database in marketing promotion, content dissemination and other business scenarios. The demand for accurate recommendations.

[0023] (4)Flexible adaptation to diverse business scenarios and object attributes

[0024] The object matching method based on attribute alignment and feature fusion of the present invention, its core process revolves around the processing of attribute description text data and the flexible extraction and fusion of features, and does not rely on specific data types that must be numerical and other restrictions. As long as the attribute information can be reflected through text description, it can be included in the consideration of matching degree calculation. This enables the method to adapt well to the changes in the attributes of different types of objects in different business scenarios. Whether it is the emergence of emerging creative style attributes of Internet celebrities or the launch of new product positioning attributes by cooperative brands, they can be integrated into the feature portrait construction and matching calculation process through corresponding text descriptions, showing strong flexibility and scalability, and providing a strong technical guarantee for long-term, continuous and accurate object matching recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0026] Figure 1 A flowchart of an object matching method based on attribute alignment and feature fusion provided in an embodiment of the present invention.

[0027] Figure 2 A schematic diagram of an object matching device based on attribute alignment and feature fusion provided by an embodiment of the present invention.

[0028] Figure 3 A structural schematic diagram of an electronic device is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be described clearly and in detail below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in the field based on the embodiments in the embodiments of the present invention should fall within the scope of protection of the embodiments of the present invention.

[0030] It should be understood that the terms "first", "second", and "third" in the claims, specifications, and drawings of the present disclosure are used to distinguish different objects rather than to describe a specific order. The terms "include" and "comprise" used in the specifications and claims of the present disclosure indicate the presence of the described features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their collections.

[0031] It should also be understood that the terms used in this disclosure are only for the purpose of describing specific embodiments and are not intended to limit the disclosure. As used in this disclosure and claims, the singular forms of "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should also be further understood that the term "and / or" used in this disclosure and claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations.

[0032] Figure 1 A flow chart of an object matching method based on attribute alignment and feature fusion provided by an embodiment of the present invention. Figure 1 As shown, it includes:

[0033] Acquire attribute description text data of the first target object, and perform feature extraction on the text data to obtain a plurality of first attribute description feature phrases;

[0034] Using a plurality of the first attribute description feature phrases to match with a set attribute classification table to cluster the first attribute description feature phrases to obtain a plurality of first dimension attribute feature sequences describing the first target object;

[0035] Vectorizing each first-dimensional attribute feature sequence to obtain a corresponding first attribute feature vector to form a feature profile of the first target object;

[0036] Acquire attribute description text data of the second target object, and perform feature extraction on the text data to obtain a plurality of second attribute description feature phrases;

[0037] Aligning a plurality of the second attribute description feature phrases with a plurality of the first attribute description feature phrases to cluster the second attribute description feature phrases to obtain a plurality of second dimension attribute feature sequences describing the second target object;

[0038] Vectorizing each second-dimensional attribute feature sequence to obtain a corresponding second attribute feature vector to form a feature profile of the second target object;

[0039] Based on the feature portrait of the first target object and the feature portrait of the second target object, the matching degree between the first target object and the second target object is calculated and compared with the set matching degree threshold to determine whether the first target object matches the second target object.

[0040] In the embodiments of the present application, the object matching method based on attribute alignment and feature fusion addresses the limitations of existing recommendation algorithms in the background technology, and can effectively improve the matching accuracy, comprehensiveness and adaptability to different business scenarios in the Internet celebrity library recommendation business, thereby helping to achieve more efficient and accurate recommendation and utilization of Internet celebrity resources.

[0041] Optionally, the attribute description text data of the first target object or the attribute description text data of the second target object is taken as a processing object, and the following steps are performed to perform feature extraction on it to obtain a result of feature extraction, wherein the result of feature extraction is a plurality of first attribute description feature phrases, or a plurality of second attribute description feature phrases:

[0042] Cleaning the text data corresponding to the processing object, removing the noise data therein, and obtaining the text to be segmented;

[0043] Based on the set semantic logic, the text to be segmented is segmented to obtain a number of words, and a business domain knowledge dictionary to be screened based on the object to be processed is obtained, and candidate attribute feature words are screened for the word set to be screened, and each word within the scope of the business domain knowledge dictionary is extracted as a candidate attribute feature word, so that all candidate attribute feature words constitute a candidate attribute feature word set;

[0044] Counting the word frequency and inverse document frequency of each candidate attribute feature word in the candidate attribute feature word set in the text data corresponding to the processing object;

[0045] Calculating the contribution to the attribute description of the target object according to the word frequency and the inverse document frequency of each candidate attribute feature word, the target object being the first target object or the second target object;

[0046] Screening out all candidate attribute feature words whose contribution is greater than a set contribution threshold, and taking each screened candidate attribute feature word as a valid attribute feature word, so that all valid attribute feature words form a valid attribute feature word set;

[0047] Based on the semantic dependency driving relationship between the semantic logic and the business domain knowledge dictionary, the effective attribute feature words in the effective attribute feature word set are relationship-constructed to obtain a plurality of effective attribute feature word groups.

[0048] To this end, the above feature extraction technology process has the following technical benefits:

[0049] 1. Improve data quality to ensure the accuracy of feature base

[0050] Cleaning the text data corresponding to the processing object can effectively remove the noise data, such as redundant punctuation, meaningless spaces, spelling errors, or some interfering text that is irrelevant to the description of the target object's attributes. Taking the description text of Internet celebrities as an example, there may be some randomly entered typos or redundant emoticons. After removing these noises, the text to be segmented is obtained, making the data for subsequent processing purer and more standardized, laying a solid foundation for accurately extracting valuable features, avoiding the extraction of erroneous or invalid features due to the interference of noise data, and thus ensuring that the feature extraction results can truly reflect the attributes of the target object.

[0051] Use the applicable business domain knowledge dictionary to filter the word set to be screened, and only extract the words within the scope of the dictionary as candidate attribute feature words. In the scenario where Internet celebrities match their partners, for example, in the beauty field where Internet celebrities are located, the dictionary will cover professional related words such as "beauty tutorials", "makeup tips", "skin care product recommendations", etc. This screening method can accurately focus on words closely related to the business, excluding those words that appear in the text but have no substantive meaning in the business context, further improving the quality of words used to construct features in the future, and ensuring that feature extraction is carried out around content with business value.

[0052] 2. Comprehensively consider word frequency and inverse document frequency to accurately measure feature importance

[0053] By counting the word frequency and inverse document frequency of each candidate attribute feature word in the candidate attribute feature word set in the text data corresponding to the processing object, and calculating the contribution to the description of the target object attribute based on this, this method can comprehensively and scientifically measure the importance of each word in describing the target object attribute. The word frequency reflects the frequency of a certain word in the current object text, and to a certain extent reflects its importance. However, relying solely on word frequency may lead to some common but too general words with low discrimination for specific objects being overvalued. The inverse document frequency considers the prevalence of the word in the entire document set (which can be understood as the overall text of the relevant object in the business scenario). By combining the two to calculate the contribution, it is possible to more accurately screen out those words that can appear frequently, are relatively unique, and are more distinguishing in the current target object text. For example, among many Internet celebrities, a certain Internet celebrity frequently mentions the "ancient style beauty makeup" feature-related words that are different from other Internet celebrities. Through such calculations, its importance can be highlighted and screened out as an effective feature, which helps to dig out feature words that can truly reflect the personality and key attributes of the target object and improve the accuracy of feature extraction.

[0054] 3. Construct effective attribute feature phrases and mine deep semantic associations

[0055] Based on the semantic dependency-driven relationship between semantic logic and business domain knowledge dictionary, the selected effective attribute feature words are constructed to obtain several effective attribute feature phrases. This operation breaks through the limitation of relying solely on a single word to describe the attributes of the object, and can dig out the deep semantic associations between words, and more comprehensively and accurately characterize the attribute characteristics of the target object. For example, in the description of the attributes of Internet celebrities, this method can be used to combine "ancient style" and "beauty" into the "ancient style beauty" feature phrase, or "food", "store exploration" and "humorous commentary" can be combined into the "humorous commentary food exploration" feature phrase. The semantic information contained in these phrases is far richer than that of a single word, and can show the attributes of Internet celebrities in content creation style, areas of expertise, etc. from multiple dimensions and in a more detailed manner, so that the subsequent feature portraits constructed based on these feature phrases can more three-dimensionally and accurately reflect the overall picture of the target object, thereby providing strong support for more accurate calculation of the matching degree with other objects, and improving the effectiveness and reliability of the entire object matching process.

[0056] 4. Adapt to the characteristics of the business field and enhance the versatility and flexibility of the method

[0057] The entire feature extraction process is closely centered around the business domain knowledge dictionary and semantic logic applicable to the processing object, which enables this method to adapt well to different business fields and application scenarios. Whether it is the attribute description text for different objects such as Internet celebrities, brand owners, event organizers, etc. in the Internet celebrity database application scenario, or the text data of different objects in other fields such as e-commerce product recommendations and social user matching, as long as the corresponding business domain knowledge dictionary is constructed and the corresponding semantic logic is followed, the attribute features that meet the business characteristics can be effectively extracted according to this set of processes. This customized processing method based on the business field enhances the versatility and flexibility of the method, enabling it to be widely used in a variety of business scenarios that require object matching based on text attribute descriptions, meeting the diverse needs of different businesses for accurate feature extraction and object matching.

[0058] Optionally, the method further comprises:

[0059] Constructing a basic information layer, a business analysis layer, and a content feature layer. The basic information layer includes regional attribute codes, industry attribute codes, and technical data attribute codes. The business analysis layer includes market influence indicators, cooperation suitability labels, and audience portrait features. The content feature layer includes: content topic labels and segmentation attribute labels.

[0060] For the basic information layer, the business analysis layer, and the content feature layer, an index key value and a corresponding data pointer are allocated, wherein the index key value is used to quantify and encode the attribute characteristics of each layer to determine the semantic logical relationship among the basic information layer, the business analysis layer, and the content feature layer, and the data pointer is used to record the semantic logical relationship among the basic information layer, the business analysis layer, and the content feature layer;

[0061] The index key value and the corresponding data pointer are filled into the structure of the designed attribute classification table to form the attribute classification table.

[0062] To this end, the above steps for constructing the attribute classification table have the following technical benefits:

[0063] 1. The layered architecture reflects multi-dimensional features and improves comprehensiveness

[0064] By constructing the basic information layer, business analysis layer, and content feature layer, and incorporating different types of attribute features, such as the basic information layer covers regional attribute codes, industry attribute codes, and technical data attribute codes, the business analysis layer includes market influence indicators, cooperation suitability labels, and audience portrait features, and the content feature layer includes content theme labels, segmentation attribute labels, etc., all kinds of important information related to the object are comprehensively sorted out and integrated from multiple dimensions. Taking the matching scenario of Internet celebrities and cooperation objects as an example, this layered architecture not only takes into account the basic information such as the region where the Internet celebrity is located and the industry to which he belongs, but also includes key elements of the business analysis level such as its market influence and audience portrait, as well as content features such as the theme and style of content creation. It can fully characterize the characteristics of the object, avoid the one-sidedness caused by considering the characteristics of the object only from a single dimension, and provide a rich and comprehensive feature foundation for subsequent precise matching.

[0065] Different business scenarios and application requirements may have different levels of focus on attributes at each level. This layered setting can flexibly extract and utilize information from one or more layers based on the specific situation. For example, in some business scenarios that focus on content cooperation, you can focus more on the relevant attributes of the content feature layer to screen matching objects; while for cooperation scenarios that value market influence and audience fit, you can highlight the use of features at the business analysis layer. The layered architecture enables the entire method to better adapt to diverse business needs and improves its applicability and effectiveness in different application scenarios.

[0066] 2. Index key values ​​and data pointers clarify semantic logic and enhance relevance and operability

[0067] By using index key values ​​to quantify and encode the attribute characteristics of each layer, the semantic logical relationship between the basic information layer, the business analysis layer, and the content feature layer can be determined in a clear and standardized manner. For example, through specific encoding rules, different cities in the regional attributes are assigned different digital codes as index key values, and industry attributes are also encoded in a corresponding orderly manner. In this way, in subsequent processing, these codes can be used to easily determine the association and hierarchical relationship between different attribute characteristics, so that the originally complex and diverse attribute information can be organized in an orderly manner according to the established logic, which is convenient for computers to process and analyze efficiently, and enhances the logic and orderliness of the entire system when processing object attribute data.

[0068] The data pointer is used to record the semantic logical relationship of each layer. It is like a navigation tool that can quickly locate and trace the relationship between attribute features of different layers. In practical applications, when it is necessary to query or use a specific attribute of an object and analyze its relationship with other attributes, the data pointer can accurately guide the system to find the corresponding information, improving the efficiency of data query and use. At the same time, it also ensures that when performing operations such as attribute feature fusion and matching calculation, information can be integrated based on accurate logical relationships to avoid error processing caused by relationship confusion, further enhancing the reliability and accuracy of the entire object matching process.

[0069] 3. The attribute classification table integrates information for unified management and efficient use

[0070] Fill the index key value and the corresponding data pointer into the structure of the designed attribute classification table to form an attribute classification table, which realizes the structured integration of the relevant information of each layer and each attribute. All key elements related to the object attribute description and their logical relationships are concentrated in this table, making the management of data resources more orderly and convenient, avoiding the inconvenience caused by decentralized data storage and management. For example, when maintaining and updating the attribute information of a large number of Internet celebrities and partners, the attribute classification table can clearly understand the status of the attributes of each object at all levels and their relationships, which is convenient for unified operation and management.

[0071] As a standardized information carrier, the attribute classification table provides a unified basis and reference standard for a series of subsequent processes such as feature extraction, clustering, vectorization, and matching degree calculation. Whether it is the first target object or the second target object, various processing operations can be accurately carried out based on this attribute classification table to ensure that different objects are analyzed and matched under the same rules and framework, ensuring the consistency and standardization of the entire object matching method, helping to improve the accuracy and comparability of the final matching results, and making the entire matching process more scientific and efficient.

[0072] Optionally, the using of a plurality of the first attribute description feature phrases to match a set attribute classification table to cluster the first attribute description feature phrases to obtain a plurality of first dimension attribute feature sequences describing the first target object includes:

[0073] For each first attribute description feature phrase, keyword matching is performed between it and each attribute category in the attribute classification table, and the semantic similarity between each first attribute description feature phrase and the content topic label corresponding to the matched attribute category in the matched attribute classification table is calculated using the trained semantic model;

[0074] If the similarity is greater than a set similarity threshold, each first attribute description feature phrase is assigned to the matched attribute category;

[0075] Clustering is performed on a plurality of first attribute description feature phrases assigned to each attribute category to obtain a plurality of first dimension attribute feature sequences describing the first target object.

[0076] To this end, the above step of using the attribute classification table to match and cluster the first attribute description feature phrases to obtain the first dimension attribute feature sequence has the following technical advantages:

[0077] 1. Use semantic similarity to achieve precise matching and improve classification accuracy

[0078] When matching the first attribute description feature phrase with the attribute category in the attribute classification table, it is not simply based on the surface matching of keywords, but further uses the trained semantic model to calculate its semantic similarity with the corresponding content theme tag. This means that it is possible to deeply explore the degree of association between words at the semantic level, surpassing the limitations of simple text matching. For example, for the attribute description feature phrase "Fun Food Sharing" of the Internet celebrity, when matching it with the content theme tag "Food Experience Introduction" of the food category in the attribute classification table, the semantic model can accurately determine that although the two are expressed differently, they are highly related in semantics, thereby achieving a more accurate match. This judgment method based on semantic similarity can better capture the true association between feature phrases and attribute categories, avoid incorrect matching due to differences in word expressions, and significantly improve the accuracy of assigning feature phrases to appropriate attribute categories.

[0079] By setting a similarity threshold to decide whether to assign a feature phrase to the matched attribute category, a quality control checkpoint is added to the matching process. The assignment is only made when the semantic similarity is greater than the threshold, which effectively filters out phrases that have a certain semantic association but a weak degree of association and may interfere with subsequent clustering and feature sequence construction. For example, in some complex business scenarios, there may be some words that have a certain semantic connection with multiple attribute categories, but they should only be classified into a certain category when they have a high similarity with a certain category and meet the actual business requirements. The setting of the similarity threshold ensures the rationality of this assignment and ensures that the feature phrases entering the subsequent clustering stage have a strong and reasonable semantic association with the corresponding attribute category, laying the foundation for accurately constructing an attribute feature sequence that reflects the characteristics of the target object.

[0080] 2. Clustering and integrating characteristic phrases to mine multi-dimensional attribute features

[0081] Clustering multiple first attribute description feature phrases assigned to each attribute category can integrate phrases with similar attribute characteristics and dig out more in-depth and comprehensive dimensional attribute features from scattered phrases. For example, in the description of Internet celebrity attributes, under the food category, there may be multiple feature phrases such as "street snack recommendations", "home cooking sharing", "food exploration in various places", etc. Through clustering, they can be integrated into a first dimension attribute feature sequence about Internet celebrities' related activities and styles in the field of food. The information contained in this sequence is far richer than that of a single phrase, and it can more comprehensively and meticulously portray the attribute characteristics of Internet celebrities in the dimension of food, which helps to fully describe the first target object from multiple different dimensions (corresponding to different attribute categories), and provide strong multi-dimensional feature support for the subsequent construction of a complete and accurate feature portrait.

[0082] Through such clustering operations, multiple first-dimensional attribute feature sequences are obtained, which can clearly present the multi-faceted and hierarchical characteristics of the first target object. Different dimensional attribute feature sequences correspond to different business attribute levels, such as regional dimension, industry dimension, content style dimension, etc. They together constitute a three-dimensional description of the target object, so that whether it is subsequent feature fusion or matching calculation with other objects, the characteristics of the object at all levels and the relationship between them can be fully considered, avoiding the one-sidedness of a single-dimensional description, and more in line with the reality that objects in actual business have complex and diverse attributes, thereby improving the comprehensiveness and accuracy of the entire object matching method in grasping the characteristics of the target object.

[0083] 3. Adapt to business logic and semantic understanding to enhance the practicality of the method

[0084] The entire process revolves around the attribute classification table, which is built based on business domain knowledge and object matching requirements. It covers different attribute categories and key information such as corresponding content topic tags, reflecting the logical relationship in the business scenario. By matching and clustering feature phrases according to the structure of the attribute classification table, the entire operation is closely aligned with the business logic and meets the needs of different object attribute classification and feature integration in actual business. For example, in the e-commerce influencer and product matching scenario, the attribute classification division of products and the corresponding topic tag settings in the attribute classification table are determined based on business processes such as product sales and promotion. The attribute description feature phrases of influencers are processed according to such rules, which can better connect and match with product attributes, enhancing the adaptability and practicality of the method in actual business applications.

[0085] The semantic similarity calculation and subsequent clustering operations based on the semantic model fully consider the semantic understanding of natural language, which is more in line with the way people describe and understand object attributes in actual business. Compared with the method based on rules or simple statistics, this semantic-based processing can better handle the diversity and flexibility of language expression, accurately extract feature information that conforms to business semantics, so that the first-dimensional attribute feature sequence finally constructed can truly and effectively reflect the key attributes of the first target object in the business scenario, providing a more reliable and practical feature basis for achieving accurate object matching, which helps to improve the effectiveness and operability of the entire object matching process in the real business environment.

[0086] Optionally, the first dimensional attribute feature sequence or the second dimensional attribute feature sequence is taken as a processing object, and the following steps are performed to vectorize it to obtain a vectorization result to form a feature portrait result, and the feature portrait result is a feature portrait of the first target object or a feature portrait of the second target object, and the vectorization result is a first attribute feature vector, or a second attribute feature vector:

[0087] According to the depth of the basic information layer, the business analysis layer, and the content feature layer, different Embedding layers are constructed respectively, and each Embedding layer has a learnable weight matrix;

[0088] Find the index key value of each attribute description feature phrase in the processing object, and vectorize it through the weight matrix of the corresponding Embedding layer to obtain a vectorized result;

[0089] Performing principal component analysis on the vectorization result to project the vectorization result into a principal component space whose spatial dimension is smaller than that of the weight matrix, and obtaining principal component feature vectors corresponding to the basic information layer, the business analysis layer, and the content feature layer;

[0090] The principal component feature vectors from the basic information layer, the business analysis layer, and the content feature layer are weightedly fused to obtain a weighted fusion vector, which is semantically annotated to generate the feature portrait.

[0091] To this end, the above steps of vectorizing the dimensional attribute feature sequence to form a feature profile have the following technical benefits:

[0092] 1. Build the Embedding layer based on layers to adapt to multi-dimensional features

[0093] Different Embedding layers are constructed according to the depth of the basic information layer, business analysis layer, and content feature layer. This hierarchical construction method can fully consider the uniqueness and differences of attribute features at different levels. For example, the relatively basic and objective information such as the regional attribute coding and industry attribute coding of the basic information layer, and the market influence indicators and audience portrait characteristics of the business analysis layer, which involve more business analysis information, and the content theme labels and segmentation attribute labels of the content feature layer, which focus on the characteristics of content creation, each have different semantics and data distribution characteristics. By constructing an independent Embedding layer for each layer, different learning mechanisms and parameter settings can be used to vectorize the feature phrases of each layer in a targeted manner, so that the vectorization results can better capture and retain the key information of each layer, and improve the ability to represent complex attribute features of different dimensions.

[0094] The importance and expression of attribute information in each layer may vary in different business scenarios, and the hierarchical Embedding layer setting can flexibly adapt to this diversity. For example, in some scenarios where influencers and brands cooperate, they may pay more attention to the characteristics of the business analysis layer to judge the degree of matching. At this time, the corresponding Embedding layer can highlight the vectorization effect of the characteristics of this layer by adjusting the weights, etc.; in other collaborations that focus on content creation collaboration, the Embedding layer of the content feature layer can play a more critical role. This flexibility allows the method to adapt to business data with different focuses and characteristics, enhancing the overall versatility and adaptability.

[0095] 2. Use index key values ​​and weight matrix vectorization to accurately represent features

[0096] The index key values ​​of each attribute description feature phrase in the processing object are searched for vectorization, with the help of the index key value system determined when the attribute classification table was previously constructed. These index key values ​​are the result of quantized encoding of attribute features at each layer. They clearly identify the category and logical relationship to which each feature phrase belongs. Through this precise positioning, it can be ensured that the corresponding feature phrase is accurately mapped to the appropriate Embedding layer and converted using the correct weight matrix during the vectorization process. This avoids confusion and incorrect mapping during the vectorization process, and ensures that the vectorization results can truly and accurately reflect the semantics of each attribute feature phrase and its position in the entire object attribute system, providing a reliable vector foundation for the subsequent construction of high-quality feature portraits.

[0097] Each Embedding layer has a learnable weight matrix. During the vectorization process, the weight matrix can transform feature phrases into appropriate vector representations by continuously learning patterns and associations in the data, so that semantically similar or related features in the vector space can reflect corresponding associations in terms of distance, direction, etc. For example, for different Internet celebrity-related feature phrases in the same food field, after learning and vectorizing the weight matrix of the Embedding layer, their positions in the vector space will be relatively close, thereby effectively using the characteristics of the vector space to represent the intrinsic connection between attribute features, improving the vectorization's ability to express features, and facilitating more accurate analysis and use of these features for object matching in the future.

[0098] 3. Principal component analysis reduces dimension, highlights key information and reduces redundancy

[0099] Performing principal component analysis on the vectorized results and projecting them into the principal component space whose spatial dimension is smaller than the weight matrix dimension can achieve data dimensionality reduction. In actual business scenarios, the original vectorized results may have high dimensions and contain some redundant information or noise, which will not only increase the complexity of subsequent calculations, but also may cover up the truly critical feature information. Through principal component analysis, we can screen out those principal component dimensions with large variance and high contribution to the overall structure of the data, and project the vectors into the principal component space composed of these key dimensions, highlighting the most important feature information and reducing the interference of irrelevant or secondary information. For example, among the many Internet celebrity attribute feature vectors, principal component analysis can focus on those dimensions that best reflect the unique style, market influence and other key aspects of the Internet celebrity, making subsequent operations based on these reduced-dimensional principal component feature vectors more efficient, and also better able to grasp the core features of the object for matching analysis.

[0100] While reducing the dimensionality, principal component analysis will retain the main feature relationships in the original data as much as possible. This means that although the dimension is reduced, the relative relationship between vectors and the key semantic associations contained in each dimension are largely preserved, and important information for judging the similarity or difference between objects will not be lost due to dimensionality reduction. In this way, when constructing feature portraits and performing matching calculations in the subsequent process, reasonable and accurate operations can still be performed based on these principal component feature vectors that retain the core feature relationships, ensuring the effectiveness and reliability of the entire object matching process.

[0101] 4. Weighted fusion and semantic annotation, integrating multi-level features and enhancing interpretability

[0102] The principal component feature vectors from different layers (basic information layer, business analysis layer, content feature layer) are weighted and fused, fully considering the different importance of features at each level to the overall description of the object. In actual business, the weights of attribute features at different layers may be different when judging the matching degree of an object. For example, for some high-end brand cooperation, the market influence features of the business analysis layer may have a higher weight; while for creative content cooperation, the relevant attributes of the content feature layer have a greater weight. By setting weights reasonably and fusing them, the key information of each layer can be organically integrated to form a vector representation that comprehensively reflects the multi-dimensional characteristics of the object, avoiding the one-sidedness of relying solely on a certain layer of features for matching, so that the feature portrait can more comprehensively and accurately portray the object, and provide a richer and more effective feature basis for precise matching.

[0103] When generating feature portraits, the weighted fusion vectors are semantically annotated, which gives the feature portraits clear semantic meaning and enhances their interpretability. In the subsequent object matching process and result analysis, people can intuitively understand the business meaning represented by each part of the feature portrait based on the semantic annotations, and understand which attribute features affect the judgment of the matching degree. For example, marking a dimension to represent the "regional influence weight" of the Internet celebrity, and another dimension to represent the "degree of uniqueness of the content style", etc., will help business personnel better understand and use feature portraits to make decisions, and also make the entire object matching method more transparent, easy to understand and control in actual business applications, improving the practicality and acceptability of the method.

[0104] Optionally, different Embedding layers are constructed respectively according to the depth of the basic information layer, the business analysis layer, and the content feature layer, and each Embedding layer has a learnable weight matrix, including:

[0105] Assign identifiers to the Embedding layers corresponding to the basic information layer, business analysis layer, and content feature layer, which are E base 、E business 、E content

[0106] For the Embedding layer corresponding to the basic information layer, business analysis layer, and content feature layer, the number of different values ​​of the input discrete features is n respectively. base 、n business 、n content , set the embedding space dimension to d;

[0107] The weight matrices of the Embedding layers corresponding to the basic information layer, business analysis layer, and content feature layer are configured as follows:

[0108] For the weight matrix of the basic information layer Embedding layer n base A matrix with d rows (corresponding to different basic information feature values) and d columns. Each element W in the matrix base [i, j] represents the weight parameter of mapping the i-th feature value in the basic information layer to the j-th dimension in the embedding space;

[0109] For the business analysis layer Embedding layer weight matrix Similarly, for n business A matrix of d rows and d columns, where the element W business [k, l] represents the weight of the k-th feature value in the embedding space l-th dimension of the business analysis layer;

[0110] For the content feature layer Embedding layer weight matrix Its function is to map various feature labels of the content feature layer to the embedding space with a dimension of n content rows, d columns, element W content [m, n] represents the weight of the mth feature value of the business analysis layer in the nth dimension of the embedding space.

[0111] To this end, the above steps on configuring weight matrices of different Embedding layers have the following technical benefits:

[0112] 1. Identification allocation facilitates differentiation and management, improving operational clarity

[0113] Different identifiers are assigned to the Embedding layers corresponding to the basic information layer, business analysis layer, and content feature layer. This approach allows the Embedding layers of each layer to be clearly distinguished in the entire object matching system. In subsequent feature processing, vectorization, and other related operations, these clear identifiers make it easy to perform targeted calls, parameter adjustments, or result analysis on the Embedding layers of different layers. For example, when you need to view the vectorization effect of the basic information layer or optimize its weight matrix, you can quickly locate the corresponding Embedding layer with the help of the corresponding identifier, avoiding confusion between different layers, improving the operational clarity and management efficiency of the entire system when processing multi-level features, and helping to carry out related work based on the features of each layer in a more orderly manner.

[0114] 2. Adapt data characteristics based on the number of feature values ​​and the dimension of the embedding space

[0115] For each Embedding layer, the number of rows of the weight matrix is ​​determined according to the different number of values ​​of the discrete features input, which fully considers the characteristics of the diversity of values ​​of the features contained in different layers. The attribute features of the basic information layer, business analysis layer and content feature layer themselves are often quite different in nature and value range. For example, the geographical attributes of the basic information layer may have only a limited number of typical values ​​(such as different city names, etc.), while the content theme tags of the content feature layer may have richer and more diverse values ​​(such as various segmented creative themes). By determining the number of rows of the weight matrix according to the number of their respective values, it can ensure that the scale of the weight matrix matches the feature richness of the corresponding layer, so that in the vectorization process, the mapping weights in the embedding space can be reasonably allocated for each different feature value, and the discrete features of each layer can be more accurately converted into appropriate vector representations, effectively adapting to the characteristic differences of data at different layers.

[0116] Setting a unified embedding space dimension provides a standardized vector mapping space for each Embedding layer. Although the features of each layer are different, vectorization in this common embedding space facilitates the subsequent integration, comparison, and analysis of vectors from different layers based on a unified spatial structure. For example, when performing subsequent steps such as principal component analysis or weighted fusion, vectors obtained in the embedding space of the same dimension are easier to process collaboratively, avoiding the complex conversion and coordination problems caused by inconsistent dimensions of vector spaces at each layer, and enhancing the standardization and consistency of multi-level feature processing in the entire object matching process, so that features at different layers can play a role in a relatively unified framework, and jointly serve to build accurate feature portraits and perform effective object matching.

[0117] 3. Clarify the meaning of weight matrix elements to enhance interpretability and controllability

[0118] The meaning of the elements in the weight matrix of each Embedding layer is explained in detail, that is, it is clearly pointed out that each element in the matrix corresponds to the weight parameter of a feature value in the corresponding layer mapped to a specific dimension of the embedding space. Such a clear definition enables technical personnel or business personnel to intuitively understand the specific meaning of each weight value when viewing and analyzing the weight matrix, and know how a feature value affects its vector representation in the embedding space during the vectorization process. For example, in the weight matrix of the Embedding layer of the basic information layer, a certain element represents the weight of the value "Beijing" in the regional attribute in a certain dimension of the embedding space. By understanding the meaning of these weights, we can better grasp the mapping relationship between features and vectors, and enhance the interpretability of the entire vectorization process and the weight matrix itself.

[0119] Since the meaning of the weight matrix elements is clear, in practical applications, if it is found that the vector representation of some features of a certain layer after vectorization does not meet expectations, or if it is hoped to strengthen or weaken the performance of some features in the embedding space, the corresponding weight parameters can be adjusted in a targeted manner according to the specific situation. This controllability makes it possible to optimize the mapping relationship from features to the embedding space according to business needs and actual data performance, further improving the quality of vectorization, so that the final vector can more accurately reflect the actual situation of the features of each layer, providing a more reliable basic guarantee for the subsequent construction of high-quality feature portraits and accurate object matching, and improving the flexibility and adaptability of the entire object matching method to deal with different business scenarios and data characteristics.

[0120] Optionally, searching for the index key value of each attribute description feature phrase in the processing object and vectorizing it through the weight matrix of the corresponding Embedding layer to obtain a vectorized result includes:

[0121] Assume that the number of attribute description feature phrases contained in the processing object is N, and for each attribute description feature phrase, it is mapped to the corresponding index key value. The corresponding index key value set configuration identifiers are I in the basic information layer, the business analysis layer, and the content feature layer respectively. base ,I business ,I content

[0122] For the basic information layer, for the i-th attribute description feature phrase (1≤i≤N), its corresponding index key value is idx base，i ∈I base , through the basic information layer Embedding layer weight matrix W base Vectorize to get vector v base，i , take the weight matrix W base The index is idx base，i The row vector of v is used as the vectorization result corresponding to the feature phrase, the vector dimension is d, and the calculation formula is: base，i =W base [idx base，s ,:]

[0123] For the business analysis layer, the jth attribute describes a characteristic phrase (1≤j≤N), and its index key value is idx business，j ∈I business , through the weight matrix W of the business analysis layer Embedding layer business The vector v obtained by vectorization business，j The calculation formula is: business，j =W business [idx business，j , :

[0124] For the kth attribute description feature phrase (1≤k≤N) of the content feature layer, the index key value is idx content，k ∈I content , through the content feature layer Embedding layer weight matrix W content Vectorization to get vector v content，k The calculation formula is: content，k =W content [dx content，k ,:]

[0125] The vectorization result vector set of the basic information layer is recorded as {v base，1 , v base，2 , …, v base，N}, business analysis layer {v business，1 , v business，2， …, v business，N}, {v content，1 , v content，2 , …, v content，N}.

[0126] To this end, the above steps of finding the index key value and vectorizing it through the corresponding Embedding layer weight matrix have the following technical benefits:

[0127] 1. The index key value set configuration identifier clearly identifies the ownership and improves processing accuracy

[0128] Different identifiers are configured for the index key value sets corresponding to each attribute description feature phrase in the processing object at the basic information layer, business analysis layer, and content feature layer. This can clearly determine the attribution of each feature phrase at different levels. In complex object attribute description scenarios, there are a large number of feature phrases of different types and sources. Through these clear identifiers, they can be quickly and accurately classified into the corresponding layers, avoiding incorrect mapping or processing caused by confusion about the attribution of feature phrases during the vectorization process. For example, when faced with many Internet celebrity attribute description feature phrases, it can accurately distinguish which ones belong to the basic information layer (such as region-related phrases), which ones belong to the business analysis layer (such as market influence-related phrases), and which ones belong to the content feature layer (such as content theme-related phrases) based on the identifiers, thereby ensuring that subsequent vectorization operations are based on the correct layer and the corresponding weight matrix, effectively improving the accuracy and orderliness of the entire vectorization process.

[0129] 2. Standardize vectorization based on index key values ​​and weight matrices to ensure mapping rationality

[0130] By mapping each attribute description feature phrase to the corresponding index key value, and then searching for the corresponding row vector in the weight matrix of each layer of Embedding layer as the vectorization result based on the index key value, an accurate and standardized vectorization method is achieved. The index key value is the embodiment of the quantized encoding of the attribute features of each layer. It has a clear correspondence with the feature phrase, and it can be used to accurately find the row in the weight matrix where the weight parameter specially configured for the feature phrase is located. This positioning method based on the index key value makes the vectorization process strictly follow the previously constructed attribute classification and encoding system, ensuring that each feature phrase can be mapped to the embedding space according to the established rules and logic, avoiding arbitrary or erroneous vector assignment, and making the conversion process from feature phrase to vector more reasonable and scientific.

[0131] The weight matrix of each layer of the Embedding layer is designed and has learnability. The vectorization results of feature phrases are generated through the row vectors of the weight matrix, which can make full use of the learning results of the mapping relationship between features and vectors contained in the weight matrix. When constructing the weight matrix, factors such as the number of feature values ​​in each layer and the dimension of the embedding space have been considered. The elements inside it represent the weight parameters of different feature values ​​in each dimension of the embedding space. Therefore, when the corresponding row vector is taken as the vectorization result based on the index key value, these carefully configured and learned weights are actually used to effectively transform the feature phrase into a vector representation that can reasonably represent its semantics and feature information in the embedding space, so that the vector can accurately reflect the characteristics of the feature phrase in the corresponding layer and its relationship with other features, providing a high-quality vector foundation for subsequent feature analysis and object matching operations.

[0132] 3. Record vectorization results in layers to facilitate subsequent layered processing and comprehensive analysis

[0133] The vectorization results of the basic information layer, business analysis layer, and content feature layer are recorded as different vector sets (\(V_{base}\), \(V_{business}\), \(V_{content}\)). This hierarchical recording method facilitates subsequent targeted processing operations and comprehensive analysis based on the characteristics of each layer. For example, when performing principal component analysis, each vector set can be independently reduced in dimension according to the business semantics and data characteristics of its layer, and the most critical principal component feature vectors of each layer can be mined; when performing weighted fusion, it can also be clearly based on the vector sets of each layer and their corresponding importance weights for reasonable fusion, fully considering the contribution of different layers to the overall object feature description. At the same time, hierarchical recording facilitates the separate evaluation and comparison of the vectorization effects of each layer, helps to promptly discover possible problems in the vectorization process of a certain layer and make optimization adjustments, ensures the efficiency and accuracy of the entire feature portrait construction based on multi-level features and object matching process, and is more in line with the needs of considering and comprehensively utilizing different dimensional attributes of objects in actual business.

[0134] Optionally, performing principal component analysis on the vectorization result to project the vectorization result into a principal component space whose spatial dimension is smaller than that of the weight matrix to obtain principal component feature vectors corresponding to the basic information layer, the business analysis layer, and the content feature layer includes:

[0135] The vectorization results of the basic information layer are organized into a vectorization result matrix by columns. Each column is a vectorized vector, with a total of N columns. The vectorization results of the business analysis layer form a vectorization result matrix The vectorization results of the content feature layer form a vectorization result matrix Calculate the mean of each column of the vectorized result matrix corresponding to the basic information layer, where for the basic information layer, the mean of each column of the vectorized result matrix μ base The calculation formula is: μ base [l] represents the mean value of the vectorization result of the basic information layer in the lth dimension; then the matrix X base Perform centralization to obtain a new matrix The calculation formula is: Calculate the mean of each column of the vectorized result matrix corresponding to the business analysis layer, where for the business analysis layer, the mean of each column of the vectorized result matrix μ business The calculation formula is: μ businness [l] represents the mean of the vectorized results of the business analysis layer in the lth dimension; then the matrix X business Perform centralization to obtain a new matrix The calculation formula is: Calculate the mean of each column of the vectorized result matrix corresponding to the content feature layer, where for the content feature layer, the mean of each column of the vectorized result matrix μ content The calculation formula is: μ content [m] represents the mean value of the vectorization result of the content feature layer in the mth dimension; then the matrix X content Perform centralization to obtain a new matrix The calculation formula is: Calculate the covariance matrix C of the vectorized result of the basic information layer base , the calculation formula is: yes The transposed matrix, covariance matrix Characterizing the variance and covariance relationship between dimensions in the vectorization result of the basic information layer;

[0140] Calculate the covariance matrix C of the vectorized results of the business analysis layer and the content feature layer business , C content , the calculation formula is: yes The transposed matrix, covariance matrix Characterize the variance and covariance relationship between the dimensions in the vectorization results of the business analysis layer; yes The transposed matrix, covariance matrix Characterize the variance and covariance relationship between dimensions in the vectorization result of the content feature layer;

[0141] The covariance matrix C corresponding to the basic information layer is base Perform eigenvalue decomposition to obtain the corresponding eigenvalues ​​and eigenvectors, and use the eigenvalues ​​in descending order λ base，1 ,λ base，2 , …, λ base，d Sort the eigenvector u base，1 ,u base，2 ,u base，d :C base u base，p =λ base，p u base，p , p = 1, 2, ..., d

[0142] The covariance matrix C corresponding to the business analysis layer bisinessPerform eigenvalue decomposition to obtain the corresponding eigenvalues ​​and eigenvectors, and use the eigenvalues ​​in descending order λ business，1 ,λ business，2 , …, λ business，d Sort the eigenvector u business，1 ,u business，2 ,…,u business，d :C business u business，p =λ business，p u business，p , p = 1, 2, ..., d

[0143] The covariance matrix C corresponding to the content feature layer is content Perform eigenvalue decomposition to obtain the corresponding eigenvalues ​​and eigenvectors, and use the eigenvalues ​​in descending order λ content，1 ,λ content，1 , …, λ content，d Sort the eigenvector u content，1 ,u content，2 ,…,u content，d :C base u base，p =λ base，p u base，p , p = 1, 2, ..., d;

[0144] For the basic information layer, the centralized matrix Projected to the principal component space, the principal component eigenvector matrix is ​​obtained The calculation formula is: Among them U base is selected by base A matrix of eigenvectors: Its dimension is d×r base ;

[0145] For the business analysis layer, select the front r business The matrix of eigenvectors The dimension is d×r business , after projection, we get the principal component eigenvector matrix The calculation formula is:

[0146] For the content feature layer, select the first content The matrix U consists of eigenvectors content : The dimension is d×r content , after projection, we get the principal component eigenvector matrix The calculation formula is:

[0147] Optionally, performing weighted fusion on the principal component feature vectors from the basic information layer, the business analysis layer, and the content feature layer to obtain a weighted fusion vector and semantically annotating the weighted fusion vector to generate the feature portrait includes:

[0148] The weights of the principal component feature vectors of the basic information layer, the business analysis layer, and the content feature layer are respectively w base 、w business 、w content , w base +w busincss +w content =1;

[0149] The corresponding principal component eigenvectors are taken out from the principal component eigenvector matrix of each layer and expressed in the form of column vectors. The corresponding basic information layer is The business analysis layer corresponds to The content feature layer corresponds to Perform weighted fusion to obtain the weighted fusion vector v combined，m , and the calculation formulas for each dimension are:

[0150] v combined，m [k] = w base × base，m [k base ]+w business × business，m [k business ]+w content ×zcontent,mkcontent

[0151] where r combined =r base +r business +r content , k base , k business , k content are the dimension indexes of the principal component eigenvectors of each layer. According to the value range of k, the corresponding elements are selected from the vectors of different layers for weighted summation, and finally the dimension is r. combined The weighted fusion vector v combined，m , which represents the comprehensive feature representation of the m-th object;

[0152] Combining the weighted fusion vectors of all objects together forms the final feature image matrix Each column corresponds to a weighted fusion vector of an object as a feature portrait vector representation;

[0153] According to the dimensional mapping relationship of each attribute feature in the attribute classification table corresponding to the basic information layer, the business analysis layer, and the content feature layer after the principal component analysis, the feature portrait matrix Vcombined Semantic annotation is performed so that each dimension in the feature portrait matrix has a business semantic interpretation.

[0154] To this end, the above principal component analysis (PCA) related steps have the following technical benefits:

[0155] 1. Data collation and centralized processing

[0156] The vectorization results of each layer (basic information layer, business analysis layer, content feature layer) are constructed into matrices (such as \(X_{base}\), \(X_{business}\), \(X_{content}\)). This method integrates the scattered vectors into a unified matrix structure, making the data presentation more regular and orderly. In the subsequent mean calculation, covariance calculation and other complex mathematical operations, batch processing can be based on the standard data organization form of matrix, which greatly improves the convenience and efficiency of data processing, facilitates the grasp and analysis of the data characteristics contained in the vectorization results of each layer from the overall level, and lays a good foundation for mining the inherent laws of data.

[0157] By calculating the mean of each column of the vectorized result matrix of each layer and performing centering processing (obtaining \(\overline{X}_{base}\), \(\overline{X}_{business}\), \(\overline{X}_{content}\)), the offset problem caused by the different means of the data itself in each dimension can be effectively eliminated. In actual situations, data of different dimensions may have different average levels due to differences in dimensions, value ranges, etc. If not centered, these mean differences will interfere with the judgment of the true relationship between dimensions and data change trends. After centering, the data is distributed around the origin, which is more helpful to highlight the relative fluctuations and mutual correlations between dimensions, making subsequent analysis based on the internal structure of the data more accurate and reliable, and creating favorable conditions for accurately extracting key feature information.

[0158] 2. Covariance matrix calculation

[0159] Calculating the covariance matrix of each layer (\(C_{base}\), \(C_{business}\), \(C_{content}\)) can clearly show the variance and covariance relationship between different dimensions in the vectorized results of each layer. Variance reflects the degree of discreteness of single-dimensional data and reflects the change characteristics of the dimension itself; while covariance reveals the linear correlation between different dimensions, that is, how the change of one dimension is accompanied by the change of other dimensions. For example, in the basic information layer, the covariance matrix can tell whether there is a linkage change between the regional attribute dimension and the industry attribute dimension and the degree of correlation. This enables us to gain a deep insight into the inherent structure hidden in multi-dimensional data, and to discover which dimensional combinations have a key impact on the overall data characteristics. It provides important clues for the subsequent targeted screening of principal components and the determination of the most representative feature dimensions, and helps to sort out the core feature association context from complex data sets, so as to more accurately grasp the key attribute characteristics of objects at all levels.

[0160] The relationship between dimensions reflected by the covariance matrix can guide subsequent feature selection and dimensionality reduction operations. Based on the contribution of each dimension to data changes and the strength of their correlation shown in the covariance matrix, we can reasonably judge which dimensions are redundant and which play a key supporting role in the overall data characteristics, and then decide which dimensions to retain as principal components for dimensionality reduction. This can avoid unnecessary computational complexity due to too much irrelevant or repetitive dimensional information in subsequent object matching processes, while ensuring that the retained principal components can cover the main feature information of the original data to the greatest extent, improve the scientificity and effectiveness of the entire object matching method in processing data, and make object matching based on these key features more practical and more accurate.

[0161] 3. Eigenvalue decomposition and sorting

[0162] The covariance matrix of each layer is decomposed by eigenvalues, and the eigenvectors are sorted from large to small according to the eigenvalues ​​(such as \(\lambda_{base}^i\), \(\lambda_{business}^i\), \(\lambda_{content}^i\) and the corresponding sorted eigenvectors). The size of the eigenvalue is of great significance, which represents the contribution of the principal component represented by the corresponding eigenvector in explaining the data variance. A larger eigenvalue means that the dimension pointed to by the corresponding eigenvector can capture more data changes and is the main manifestation of the overall data change trend. In this way, we can accurately select the most critical and representative few dimensions from many original dimensions to describe the overall characteristics of the data, and focus on these core feature dimensions for subsequent operations, so that in the object matching process, we can focus on the key features that best reflect the differences and similarities between objects, thereby significantly improving the accuracy and efficiency of matching judgments and avoiding misjudgments due to the interference of too many secondary features.

[0163] The first several eigenvectors are selected in order of eigenvalue size to form a projection matrix (such as \(U_{base}\), \(U_{business}\), \(U_{content}\)) to construct the principal component space. This orderly and principle-based principal component screening mechanism ensures that in the process of dimensionality reduction, the main information of the original data can be retained as much as possible while effectively removing those relatively minor dimensions that have little impact on the overall characteristics. For example, when projecting the vectorized results to the principal component space, only the most important principal component eigenvectors are selected. This not only achieves a significant simplification of the data dimension and reduces the burden of data storage and subsequent calculations, but also the obtained principal component eigenvectors can still well preserve the core feature relationship of the original data, providing a high-quality feature foundation for the subsequent construction of high-quality and more discriminative feature portraits and accurate object matching, and optimizing the data quality and processing efficiency of the entire object matching process.

[0164] 4. Projection to the principal component space

[0165] By projecting the centralized matrix into the principal component space (such as \(\overline{X}_{base}U_{base}\), \(\overline{X}_{business}U_{business}\), and \(\overline{X}_{content}U_{content}\) to obtain the principal component eigenvector matrices of each layer respectively), the goal of dimensionality reduction is achieved. The high-dimensional space where the original vectorized results are located often contains a lot of redundant information and noise, which not only takes up a lot of storage resources, but also makes the subsequent data calculation, analysis and other operations complicated and time-consuming, seriously affecting the efficiency of the entire object matching process, especially when facing large-scale object data. After being projected into the principal component space, the data is converted into a space with lower dimensions and more compactness. The main features of the original data can be effectively represented with fewer dimensions, which greatly reduces the amount of data and the computational complexity of subsequent processing, speeds up the calculation speed, and enables the entire object matching method to more efficiently cope with the matching needs of a large number of objects and meet the requirements for fast and accurate matching in actual business scenarios.

[0166] In the dimensionality reduction process, since the projection is based on the key principal components selected by eigenvalue decomposition, the principal component feature vectors finally obtained optimize the feature representation while retaining the core feature relationship of the original data. These principal component feature vectors focus more on the differences and similarities of key attributes of objects at different levels (basic information layer, business analysis layer, content feature layer), and can more accurately characterize the essential characteristics of the object. Compared with the original vectorization results, their description of object characteristics is more concise, clear and representative. In the subsequent weighted fusion of the principal component feature vectors of each layer and the construction of a complete feature portrait, these optimized feature vectors can provide more powerful support for accurately judging the matching degree between objects, thereby improving the accuracy and reliability of the entire object matching process, and better serving the needs of accurately matching different objects in actual business.

[0167] Optionally, the feature portrait matrix of the first target object is The matrix contains m feature image vectors, Represents its i-th feature image vector (i = 1, 2, ..., m), and the dimension of each vector is r 1 : The feature image matrix of the second target object is: This matrix contains n feature image vectors, Represents the jth feature image vector (j=1, 2, ..., n), and the dimension of each vector is r 2 ;

[0168] The calculating the matching degree between the first target object and the second target object based on the feature portrait of the first target object and the feature portrait of the second target object and comparing the matching degree with a set matching degree threshold to determine whether the first target object matches the second target object includes:

[0169] Based on the formula Calculate the similarity between the i-th feature portrait vector of the first target object and the j-th feature portrait vector of the second target object to obtain an m×n similarity matrix SimilarityMatrix, represents the k-th dimension element of the i-th feature portrait vector of the first target object, The k-th dimension element representing the j-th feature portrait vector of the second target object;

[0170] Based on the formula The overall similarity between the first target object and the second target object is calculated and compared with a set matching threshold to determine whether the first target object matches the second target object.

[0171] To this end, the above steps of calculating the matching degree between the first target object and the second target object based on the feature portrait have the following technical benefits:

[0172] 1. Calculate similarity based on vector dimension elements

[0173] -**Comprehensively consider the relationship between features in each dimension**:

[0174] The formula (SimilarityMatrix(i, j) is used to calculate the similarity between the \(i\)th feature portrait vector of the first target object and the \(j\)th feature portrait vector of the second target object. This method fully considers the relationship between the elements of each dimension of the two vectors. When constructing the feature portrait vector, each dimension incorporates the key attribute information of the object at different levels (such as basic information, business analysis, content characteristics, etc.), and each dimension element carries specific business semantics. Covering all dimensional elements when calculating the similarity means comprehensively measuring the similarity of the two objects in the corresponding feature portrait vectors from multiple key dimensions, avoiding the one-sidedness of only focusing on some dimensions and ignoring the impact of other important attributes on matching. It can more comprehensively and meticulously capture the fit of the characteristics of the two objects in various aspects, making the calculation result of the similarity more valuable for reference and more truly reflecting the closeness of the association between objects based on multi-dimensional features.

[0175] The above formula essentially calculates similarity based on the inner product and vector modulus of the vector, and its principle utilizes the characteristics of the vector space. In the vector space, the inner product of the vector reflects the relationship between the vectors in direction and length, while the vector modulus characterizes the length property of the vector itself. By determining the similarity through the ratio of the two, we can accurately judge the relative position relationship and similarity of the two vectors in space. For example, if the feature portrait vectors of two objects are relatively consistent in direction in the vector space (meaning that the sum of the products between the elements of each dimension is relatively large, and the product of the vector modulus is also reasonably matched), then their similarity will be high, which is in good agreement with our business needs to measure whether two objects are similar from the perspective of multi-dimensional features, and provides a reliable quantitative basis based on vector space theory for the subsequent accurate judgment of whether the objects match or not.

[0176] 2. Calculate the overall matching degree

[0177] The formula OverallSimilarity is used to calculate the overall matching degree between the first target object and the second target object, which is a comprehensive consideration of the similarity between all feature portrait vectors. Because the feature portrait matrix of the first target object contains \(m\) vectors, and the feature portrait matrix of the second target object contains \(n\) vectors, by calculating and summarizing the similarities between all these vector pairs and then averaging them, it is possible to reflect the comprehensive matching of the two objects in the multi-dimensional features covered by multiple feature portrait vectors from an overall level. Different feature portrait vectors may characterize the attributes of the object from different angles, stages or aspects. The overall matching degree is obtained by combining the similarities between them, avoiding the limitation of relying solely on the similarity of a pair of vectors to judge the matching of objects, and more comprehensively reflecting the overall degree of fit between the two objects, providing a more macro and comprehensive basis for the final accurate judgment of whether they match.

[0178] The calculated overall matching degree is compared with the set matching degree threshold to determine whether the first target object matches the second target object, providing a clear and unambiguous judgment standard for object matching. In actual business scenarios, different business needs and application scenarios may have different requirements for the degree of matching. By pre-setting appropriate matching degree thresholds, it is possible to flexibly determine what kind of object combination meets business requirements and can be considered a match based on specific circumstances. This threshold-based judgment mechanism makes the entire matching process more in line with actual business logic, and can accurately screen out object pairs that meet specific matching criteria, which helps to improve the effectiveness and applicability of object matching results in actual business applications and meet diverse business needs. For example, in the matching scenario between Internet celebrities and partners, different thresholds can be set according to the requirements for the closeness of cooperation to select appropriate matching combinations.

[0179] Based on the above embodiment, the embodiment of the present application further provides an object matching method, which includes:

[0180] Obtaining attribute description data of each of the first target object and the second target object to be matched;

[0181] Performing feature extraction processing on the acquired attribute description data of each target object respectively to obtain multiple attribute feature groups corresponding to each target object;

[0182] Classify and cluster the attribute feature groups of each target object according to preset rules to form a multi-dimensional attribute feature sequence corresponding to each target object;

[0183] Vectorization operations are performed on the multi-dimensional attribute feature sequences of each target object to construct feature portraits of each target object;

[0184] Based on the constructed feature portraits of the first target object and the second target object, the degree of match between the two is calculated and compared with a preset matching degree judgment standard to determine whether the first target object and the second target object match each other.

[0185] Figure 2 Schematic diagram of an object matching device based on attribute alignment and feature fusion provided by an embodiment of the present invention. Figure 2 As shown, it includes:

[0186] A first program unit is used to obtain attribute description text data of a first target object and perform feature extraction on the text data to obtain a plurality of first attribute description feature phrases;

[0187] A second program unit is used to match a plurality of the first attribute description feature phrases with a set attribute classification table to cluster the first attribute description feature phrases to obtain a plurality of first dimension attribute feature sequences describing the first target object;

[0188] A third program unit is used to vectorize each first-dimensional attribute feature sequence to obtain a corresponding first attribute feature vector to form a feature portrait of the first target object;

[0189] The fourth program unit is used to obtain the attribute description text data of the second target object and perform feature extraction on it to obtain a plurality of second attribute description feature phrases;

[0190] A fifth program unit is used to align a plurality of the second attribute description feature phrases with a plurality of the first attribute description feature phrases to cluster the second attribute description feature phrases to obtain a plurality of second dimension attribute feature sequences describing the second target object;

[0191] A sixth program unit, configured to vectorize each second-dimensional attribute feature sequence to obtain a corresponding second attribute feature vector to form a feature portrait of the second target object;

[0192] The seventh program unit is used to calculate the matching degree between the first target object and the second target object based on the feature portrait of the first target object and the feature portrait of the second target object, and compare it with the set matching degree threshold to determine whether the first target object matches the second target object.

[0193] Above Figure 2 In the embodiments, the exemplary explanation of the technical processing process of each program unit can be found in the above Figure 1 Records of.

[0194] Figure 3 A schematic diagram of the structure of an electronic device is provided in accordance with an embodiment of the present invention. Figure 3 As shown, the electronic device includes a memory and a processor, the memory stores a computer executable program, and the processor runs the computer executable program to implement the following steps:

[0195] Acquire attribute description text data of the first target object, and perform feature extraction on the text data to obtain a plurality of first attribute description feature phrases;

[0196] Using a plurality of the first attribute description feature phrases to match with a set attribute classification table to cluster the first attribute description feature phrases to obtain a plurality of first dimension attribute feature sequences describing the first target object;

[0197] Vectorizing each first-dimensional attribute feature sequence to obtain a corresponding first attribute feature vector to form a feature profile of the first target object;

[0198] Acquire attribute description text data of the second target object, and perform feature extraction on the text data to obtain a plurality of second attribute description feature phrases;

[0199] Aligning a plurality of the second attribute description feature phrases with a plurality of the first attribute description feature phrases to cluster the second attribute description feature phrases to obtain a plurality of second dimension attribute feature sequences describing the second target object;

[0200] Vectorizing each second-dimensional attribute feature sequence to obtain a corresponding second attribute feature vector to form a feature profile of the second target object;

[0201] Based on the feature portrait of the first target object and the feature portrait of the second target object, the matching degree between the first target object and the second target object is calculated and compared with the set matching degree threshold to determine whether the first target object matches the second target object.

[0202] Above Figure 3 In the embodiment, the exemplary explanation of the technical processing process of each step can be found in the above Figure 1 Records of.

[0203] The above implementation methods are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims. The systems, devices, modules or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0204] For the convenience of description, the above device is described as being divided into various units according to their functions. Of course, when implementing the present invention, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0205] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. An object matching method based on attribute alignment and feature fusion, characterized in that: include: Acquire attribute description text data of the first target object, and perform feature extraction on the text data to obtain a plurality of first attribute description feature phrases; Using a plurality of the first attribute description feature phrases to match with a set attribute classification table to cluster the first attribute description feature phrases to obtain a plurality of first dimension attribute feature sequences describing the first target object; Vectorizing each first-dimensional attribute feature sequence to obtain a corresponding first attribute feature vector to form a feature profile of the first target object; Acquire attribute description text data of the second target object, and perform feature extraction on the text data to obtain a plurality of second attribute description feature phrases; Aligning a plurality of the second attribute description feature phrases with a plurality of the first attribute description feature phrases to cluster the second attribute description feature phrases to obtain a plurality of second dimension attribute feature sequences describing the second target object; Vectorizing each second-dimensional attribute feature sequence to obtain a corresponding second attribute feature vector to form a feature profile of the second target object; Based on the feature portrait of the first target object and the feature portrait of the second target object, the matching degree between the first target object and the second target object is calculated and compared with the set matching degree threshold to determine whether the first target object matches the second target object.

2. The object matching method based on attribute alignment and feature fusion according to claim 1, characterized in that: The attribute description text data of the first target object or the attribute description text data of the second target object is taken as a processing object, and the following steps are performed to extract features thereon to obtain a result of feature extraction, wherein the result of feature extraction is a plurality of first attribute description feature phrases or a plurality of second attribute description feature phrases: Cleaning the text data corresponding to the processing object, removing the noise data therein, and obtaining the text to be segmented; Based on the set semantic logic, the text to be segmented is segmented to obtain a number of words, and a business domain knowledge dictionary to be screened based on the object to be processed is obtained, and candidate attribute feature words are screened for the word set to be screened, and each word within the scope of the business domain knowledge dictionary is extracted as a candidate attribute feature word, so that all candidate attribute feature words constitute a candidate attribute feature word set; Counting the word frequency and inverse document frequency of each candidate attribute feature word in the candidate attribute feature word set in the text data corresponding to the processing object; Calculating the contribution to the attribute description of the target object according to the word frequency and the inverse document frequency of each candidate attribute feature word, the target object being the first target object or the second target object; Screening out all candidate attribute feature words whose contribution is greater than a set contribution threshold, and taking each screened candidate attribute feature word as a valid attribute feature word, so that all valid attribute feature words form a valid attribute feature word set; Based on the semantic dependency driving relationship between the semantic logic and the business domain knowledge dictionary, the effective attribute feature words in the effective attribute feature word set are relationship-constructed to obtain a plurality of effective attribute feature word groups.

3. The object matching method based on attribute alignment and feature fusion according to claim 1, characterized in that: The method further comprises: Constructing a basic information layer, a business analysis layer, and a content feature layer. The basic information layer includes regional attribute codes, industry attribute codes, and technical data attribute codes. The business analysis layer includes market influence indicators, cooperation suitability labels, and audience portrait features. The content feature layer includes: content topic labels and segmentation attribute labels. For the basic information layer, the business analysis layer, and the content feature layer, an index key value and a corresponding data pointer are allocated, wherein the index key value is used to quantify and encode the attribute characteristics of each layer to determine the semantic logical relationship among the basic information layer, the business analysis layer, and the content feature layer, and the data pointer is used to record the semantic logical relationship among the basic information layer, the business analysis layer, and the content feature layer; The index key value and the corresponding data pointer are filled into the structure of the designed attribute classification table to form the attribute classification table.

4. The object matching method based on attribute alignment and feature fusion according to claim 3, characterized in that: The using of a plurality of the first attribute description feature phrases to match with a set attribute classification table to cluster the first attribute description feature phrases to obtain a plurality of first dimension attribute feature sequences describing the first target object includes: For each first attribute description feature phrase, keyword matching is performed between it and each attribute category in the attribute classification table, and the semantic similarity between each first attribute description feature phrase and the content topic label corresponding to the matched attribute category in the matched attribute classification table is calculated using the trained semantic model; If the similarity is greater than a set similarity threshold, each first attribute description feature phrase is assigned to the matched attribute category; Clustering is performed on a plurality of first attribute description feature phrases assigned to each attribute category to obtain a plurality of first dimension attribute feature sequences describing the first target object.

5. The object matching method based on attribute alignment and feature fusion according to claim 4, characterized in that: The first dimensional attribute feature sequence or the second dimensional attribute feature sequence is taken as a processing object, and the following steps are performed to vectorize it to obtain a vectorization result to form a feature portrait result, wherein the feature portrait result is a feature portrait of the first target object or a feature portrait of the second target object, and the vectorization result is a first attribute feature vector or a second attribute feature vector: According to the depth of the basic information layer, the business analysis layer, and the content feature layer, different Embedding layers are constructed respectively, and each Embedding layer has a learnable weight matrix; Find the index key value of each attribute description feature phrase in the processing object, and vectorize it through the weight matrix of the corresponding Embedding layer to obtain a vectorized result; Performing principal component analysis on the vectorization result to project the vectorization result into a principal component space whose spatial dimension is smaller than that of the weight matrix, and obtaining principal component feature vectors corresponding to the basic information layer, the business analysis layer, and the content feature layer; The principal component feature vectors from the basic information layer, the business analysis layer, and the content feature layer are weightedly fused to obtain a weighted fusion vector, which is semantically annotated to generate the feature portrait.

6. The object matching method based on attribute alignment and feature fusion according to claim 5, characterized in that: According to the depth of the basic information layer, the business analysis layer, and the content feature layer, different Embedding layers are constructed respectively, and each Embedding layer has a learnable weight matrix, including: Assign identifiers to the Embedding layers corresponding to the basic information layer, business analysis layer, and content feature layer, which are E base 、E business 、E content ; For the Embedding layer corresponding to the basic information layer, business analysis layer, and content feature layer, the number of different values ​​of the input discrete features is n respectively. base 、n business 、n content , set the embedding space dimension to d; The weight matrices of the Embedding layers corresponding to the basic information layer, business analysis layer, and content feature layer are configured as follows: For the weight matrix of the basic information layer Embedding layer n base A matrix with d rows (corresponding to different basic information feature values) and d columns. Each element W in the matrix base [i,] represents the weight parameter of mapping the i-th feature value in the basic information layer to the j-th dimension in the embedding space; For the business analysis layer Embedding layer weight matrix Similarly, for n business A matrix with d rows and d columns, where the element W business [k, l] represents the weight of the k-th feature value in the embedding space l-th dimension of the business analysis layer; For the content feature layer Embedding layer weight matrix Its function is to map various feature labels of the content feature layer to the embedding space with a dimension of n content rows, d columns, element W content [m, n] represents the weight of the mth feature value of the business analysis layer in the nth dimension of the embedding space.

7. The object matching method based on attribute alignment and feature fusion according to claim 5, characterized in that: The searching for the index key value of each attribute description feature phrase in the processing object and vectorizing it through the weight matrix of the corresponding Embedding layer to obtain the vectorization result includes: Assume that the number of attribute description feature phrases contained in the processing object is N, and for each attribute description feature phrase, it is mapped to the corresponding index key value. The corresponding index key value set configuration identifiers are I in the basic information layer, the business analysis layer, and the content feature layer respectively. base ,I business ,I content ; For the basic information layer, for the i-th attribute description feature phrase (1≤i≤N), its corresponding index key value is idx base，i ∈I base , through the basic information layer Embedding layer weight matrix W base Vectorize to get vector v base，i , take the weight matrix W base The index is idx base，i The row vector of v is used as the vectorization result corresponding to the feature phrase, the vector dimension is d, and the calculation formula is: base，i =W base [idx base，i ,:] For the business analysis layer, the jth attribute describes the characteristic phrase (1≤j≤N), and its index key value is idx business，j ∈I business , through the weight matrix W of the business analysis layer Embedding layer business The vector u obtained by vectorization business，j The calculation formula is: business，j =W business [idx business，j ,:] For the kth attribute description feature phrase (1≤k≤N) of the content feature layer, the index key value is idx content，k ∈I content , through the content feature layer Embedding layer weight matrix W content Vectorization to get vector v content，k The calculation formula is: content，k =W content [idx content，k ,:] The vectorization result vector set of the basic information layer is recorded as {v base，1 , v base，2 , …, v base，N }, business analysis layer {v business，1 , v business，2 , …, v business，N }, {v content，1 , v content，2 , …, v content，N }.

8. The object matching method based on attribute alignment and feature fusion according to claim 5, characterized in that: The performing principal component analysis on the vectorization result to project the vectorization result into a principal component space whose spatial dimension is smaller than that of the weight matrix, and obtaining principal component feature vectors corresponding to the basic information layer, the business analysis layer, and the content feature layer, includes: The vectorization results of the basic information layer are organized into a vectorization result matrix by columns. Each column is a vectorized vector, with a total of N columns. The vectorization results of the business analysis layer form a vectorization result matrix The vectorization results of the content feature layer form a vectorization result matrix Calculate the mean of each column of the vectorized result matrix corresponding to the basic information layer, where for the basic information layer, the mean of each column of the vectorized result matrix μ base The calculation formula is: μ base [l] represents the mean value of the vectorization result of the basic information layer in the lth dimension; then the matrix X base Perform centralization to obtain a new matrix The calculation formula is: l=1,2,…,d,i=1,2,…,N; Calculate the mean of each column of the vectorized result matrix corresponding to the business analysis layer, where for the business analysis layer, the mean of each column of the vectorized result matrix μ business The calculation formula is: μ business [l] represents the mean of the vectorized results of the business analysis layer in the lth dimension; then the matrix X business Perform centralization to obtain a new matrix The calculation formula is: l=1,2,…,d,i=1,2,…,N; Calculate the mean of each column of the vectorized result matrix corresponding to the content feature layer, where for the content feature layer, the mean of each column of the vectorized result matrix μ content The calculation formula is: μ content [m] represents the mean value of the vectorization result of the content feature layer in the mth dimension; then the matrix X content Perform centralization to obtain a new matrix The calculation formula is: m=1,2,…,d,i=1,2,…,N; Calculate the covariance matrix C of the vectorized result of the basic information layer base , the calculation formula is: yes The transposed matrix, covariance matrix Characterizing the variance and covariance relationship between dimensions in the vectorization result of the basic information layer; Calculate the covariance matrix C of the vectorized results of the business analysis layer and the content feature layer business , C content , the calculation formula is: yes The transposed matrix, covariance matrix Characterize the variance and covariance relationship between the dimensions in the vectorization results of the business analysis layer; yes The transposed matrix, covariance matrix Characterize the variance and covariance relationship between dimensions in the vectorization result of the content feature layer; The covariance matrix C corresponding to the basic information layer is base Perform eigenvalue decomposition to obtain the corresponding eigenvalues ​​and eigenvectors, and use the eigenvalues ​​in descending order λ base，1 ,λ base，2 , …, λ base，d Sort the eigenvector u base，1 ,u base，2 ,…,u base，d :C base u base，p =λ base，p u base，p , p = 1, 2, ..., d; The covariance matrix C corresponding to the business analysis layer business Perform eigenvalue decomposition to obtain the corresponding eigenvalues ​​and eigenvectors, and use the eigenvalues ​​in descending order λ business，1 ,λ business，2 , …,λ business,d Sort the eigenvector u business，1 ,u business，2 ,…,u business，d :C business u business，p =λ business ,pu business，p , p = 1, 2, ..., d; The covariance matrix C corresponding to the content feature layer is content Perform eigenvalue decomposition to obtain the corresponding eigenvalues ​​and eigenvectors, and use the eigenvalues ​​in descending order λ content，1 ,λ content，1 , …, λ content，d Sort the eigenvector u content，1 ,u content，2 ,…,u content，d :C base u base，p =λ base，p u base，p , p = 1, 2, ..., d; For the basic information layer, the centralized matrix Projected to the principal component space, the principal component eigenvector matrix is ​​obtained The calculation formula is: Among them U base is selected by base The matrix composed of eigenvectors: U base =[ base ,1,u base , 2,…,u base , r base ], whose dimension is d×r base . For the business analysis layer, select the front r business The matrix of eigenvectors The dimension is d×r business , after projection, we get the principal component eigenvector matrix The calculation formula is: For the content feature layer, select the first content The matrix U consists of eigenvectors content : The dimension is d×r content , after projection, we get the principal component eigenvector matrix The calculation formula is:

9. The object matching method based on attribute alignment and feature fusion according to claim 5, characterized in that: The step of weighting and fusing the principal component feature vectors from the basic information layer, the business analysis layer, and the content feature layer to obtain a weighted fusion vector and semantically annotating the weighted fusion vector to generate the feature portrait includes: The weights of the principal component feature vectors of the basic information layer, the business analysis layer, and the content feature layer are respectively w base 、w business 、w content , w base +w business +w content =1; The corresponding principal component eigenvectors are taken out from the principal component eigenvector matrix of each layer and expressed in the form of column vectors. The corresponding basic information layer is The business analysis layer corresponds to The content feature layer corresponds to Perform weighted fusion to obtain the weighted fusion vector v combined，m , and the calculation formulas for each dimension are: v combined，m [k]=w base ×z base，m [k base ]+w business ×z business，m [k business ]+w content ×z content，m [k content ] where r conbined =r base +r business +r content , k base , k business , k content are the dimension indexes of the principal component eigenvectors of each layer. According to the value range of k, the corresponding elements are selected from the vectors of different layers for weighted summation, and finally the dimension is r. combined The weighted fusion vector v combined，m , which represents the comprehensive feature representation of the m-th object. Combining the weighted fusion vectors of all objects together forms the final feature image matrix The weighted fusion vector of each column corresponding to an object is represented as a feature portrait vector; According to the dimensional mapping relationship of each attribute feature in the attribute classification table corresponding to the basic information layer, the business analysis layer, and the content feature layer after the principal component analysis, the feature portrait matrix V combined Semantic annotation is performed so that each dimension in the feature portrait matrix has a business semantic interpretation.

10. The object matching method based on attribute alignment and feature fusion according to claim 5, characterized in that: The feature portrait matrix of the first target object is: The matrix contains m feature image vectors, Represents its i-th feature image vector (i=1, 2, ..., m), and the dimension of each vector is r1: The feature image matrix of the second target object is: This matrix contains n feature image vectors, represents its j-th feature portrait vector (j=1, 2, ..., n), and the dimension of each vector is r2; The calculating the matching degree between the first target object and the second target object based on the feature portrait of the first target object and the feature portrait of the second target object and comparing the matching degree with a set matching degree threshold to determine whether the first target object matches the second target object includes: Based on the formula Calculate the similarity between the i-th feature portrait vector of the first target object and the j-th feature portrait vector of the second target object to obtain an m×n similarity matrix SimilarityMatrix, represents the k-th dimension element of the i-th feature portrait vector of the first target object, The k-th dimension element representing the j-th feature portrait vector of the second target object; Based on the formula The overall similarity between the first target object and the second target object is calculated and compared with a set matching threshold to determine whether the first target object matches the second target object.

Images